Tuning tender alerts so your team actually reads them

To make tender alerts precise enough that people read them, count your precision over one month, rebuild each failing search with classification codes as its spine instead of keywords, then add disqualifiers that strip out whole categories of noise. Replay the result against notices you know were relevant before switching the old setup off.

Alert fatigue is why good pipelines rot. An inbox running at nine parts noise to one part signal stops being read within a month, and the one part goes unread with it. This takes ninety minutes and runs straight after your quarterly check of what your alerts missed: coverage says what to widen, this what to narrow.

How to measure your alert precision

Pick one complete month that ended at least thirty days ago. Count distinct notices, not emails: a digest of nine notices is nine, and the same procurement from two portals is one. Call that total N.

Then one person, in one pass, marks each notice as worth a proper look or not: it went, or should have gone, into your ten minute qualification decision, not that you bid it. Call that count U.

Precision is U divided by N. Compute it overall, then per saved search, which tells you what to rewrite, then per portal. Count duplicates separately: deduplication buys more than any keyword edit and costs no coverage.

One tuning round, same month replayed

Why keyword alerts fail in procurement

Keyword alerting works where sellers write the listings. Procurement is the opposite, and it fails in four specific ways.

• Buyers describe the problem, not the product. A council replacing the system its social workers use publishes "adult social care case management modernisation".
• Titles are administrative. Plenty read "framework agreement for the provision of services, lot 3", with the substance in the description field or only in the tender documents, which search does not read.
• Wording varies by language. eForms, mandatory for new EU procedures since 25 October 2023 under Commission Implementing Regulation (EU) 2019/1780, standardised the structured fields but not the free text, which stays in the buyer's language. Hence matching in the notice's own language.
• The same work appears under works framing or services framing. A roof replacement scoped as construction work carries different words from the same job scoped as maintenance.

How to build alerts on classification codes

Every notice carries a classification code because the buyer must pick one. Mandatory metadata beats prose.

CPV codes run across the EU and the UK: eight digits plus a check digit, a tree from broad division down to specific commodity. Watch at division or group level and let the portal include everything beneath, because buyers choose the leaf inconsistently. On TED, sectors and place of performance are hierarchical, so the parent brings the children.

For US federal work, NAICS codes and product service codes come at the notice from two angles. NAICS describes the supplier's industry and sets the small business size standard, so the same work appears under different codes at different agencies; PSC describes what is being bought. Filter SAM.gov on both. NAICS 2022 is the current version; the Census Bureau's NAICS pages carry the concordance files. UN agencies and development banks use UNSPSC, where the codes you pick at registration decide what reaches you.

How to find the codes your competitors' contracts carried

Read them off awards. Search contract award notices from the last twelve months on TED, Contracts Finder and Find a Tender and record the CPV code on every award you would have wanted. For US federal work, USAspending advanced search filters awards by NAICS and PSC and searches on recipient name, so you can list what a competitor has been paid under.

Take the union, sort by frequency, cut the tail. A code on your coverage audit miss list but not in your alert setup is the cheapest fix of the quarter.

Which disqualifier terms to add

A disqualifier earns its place when it removes a lot of alerts and nothing you wanted. Start with the structural ones, which need no sector judgement.

• Send corrigenda and change notices out of the discovery feed and to whoever owns that bid.
• Send award notices to a separate monthly review. They are competitor intelligence, not opportunities.

Then the vocabulary ones. A cyber security firm excludes "manned guarding", "door supervision" and "CCTV installation", killing physical security notices that match on the word security. A software firm excludes "toner" and "cartridge", killing consumables that match on "supply and support of". An engineering consultancy excludes "agency staff" and "temporary workers", killing staffing frameworks.

Before adding a term as a negative, run it as a positive search over the last quarter. If one notice in that list is one you would have bid, the term stays.

The keyword audit table

One row per term in your worst-scoring search. A worked example:

Anything under roughly 10% precision that carries volume is where the fatigue comes from.

How to set value and geography filters without cutting yourself off

Framework ceilings are the second trap. A framework published at its ceiling looks enormous while the lot you would bid is a fraction of it. Filter on a minimum, never a maximum, and set the floor at half your smallest viable contract.

Geography has an equivalent trap: filtering on buyer address rather than place of performance. Central purchasing bodies buy from a capital city for delivery everywhere, so a filter on buyer location deletes that work from view.

GOV.UK describes Contracts Finder as covering contracts worth over £12,000 including VAT, while Find a Tender carries the high value notices and has sat on the Central Digital Platform since the Procurement Act 2023 regime went live in February 2025. Both let an account holder save search criteria and sign up for email notifications, so re-check any query written before then.

How to test a change before you trust it

Never switch a live alert because the new query looks sensible. Replay it.

1. Build a known-good set of 20 to 40 notices from the last quarter you are certain were relevant: your bid tracker, alerts you acted on, the coverage audit miss list. The misses matter most, being the cases the old setup failed.
2. Run the new query as a historical search on each portal, bounded to that date range. TED, SAM.gov, Contracts Finder and Find a Tender all support publication date ranges.
3. Count how many come back. That is recall on a small sample, and the number that protects you. Trace every loss to the filter that dropped it.
4. Count the total returned: that is your projected volume, and tells you whether the result is readable. Then run old and new side by side for two weeks and compare what only one caught.

The replay gives recall and volume, not precision: the known-good set is not exhaustive. Precision needs a fresh month next quarter.

Who should receive which alerts

One shared inbox for everything guarantees fatigue, for a reason unrelated to volume. An alert with no owner is optional reading, and optional work loses to work with a deadline.

• One named owner per category, plus a deputy for holidays and illness.
• Immediate delivery only where deadlines are short. Everything else goes into one daily digest, timed to land before the daily portal round.
• On SAM.gov, following an individual notice pushes changes as they happen: right for a bid you are writing, wrong for discovery.
• One unrouted queue for alerts matching no category. A growing queue means your categories have drifted from the market.

The TED saved search and alert limits shape what routing is possible: 25 saved searches and 25 search alerts per account, at daily, weekly, fortnightly or monthly frequency. Worth knowing before you design a scheme needing forty.

The quarterly tuning checklist

1. Run the coverage audit first. Widen before you narrow.
2. Pull one complete month of alerts and count distinct notices.
3. Mark each as worth a proper look or not. One person, one pass.
4. Compute precision overall, per saved search and per portal.
5. Rank the failing searches by volume and fix only the biggest.
6. Fill in the keyword audit table for every term in it.
7. Pull last quarter's awards in your sector and record their codes.
8. Rewrite the search: codes as the spine, keywords at the edges, disqualifiers for known noise.
9. Check the value floor against blank-value notices, and that geography uses place of performance.
10. Replay against the known-good set, then run old and new in parallel for two weeks.
11. Switch, and log the date and the change.
12. Confirm an owner and a deputy per category, and review the unrouted queue.
13. Record precision and volume. Change at most three searches a round.

Each quarter, watch the trend in precision rather than the absolute number, the size of the unrouted queue, and whether a portal has changed its notice types or fields. That last one is the failure your own query cannot protect you from, and part of deciding which portals to watch at all.

When precision cannot be fixed by filtering

Two rounds of tuning should move the number. If it does not, and the useful notices are scattered thinly across codes rather than clustered, filtering is not your problem and more of it will only cost coverage. Four situations produce that pattern.

You sell a component of larger contracts. The notice describes a school and you supply the cladding, so no filter helps: your work is never named. The route in is award notices and the firms that win them.

Your differentiator sits in the qualification criteria rather than the subject. You win where the buyer allows consortium bids, or accepts your accreditation, or splits into small enough lots. No code encodes that.

The market is simply small. If six relevant notices exist in a quarter, precision is the wrong measure and chasing it will make you miss one of the six.

Two people disagree about what counted as useful. Until your team applies the same standard to the same notice, every precision figure you compute measures who did the marking.

Common questions

How many tender alerts a day is too many?

What matters is how many get opened, not how many arrive. If fewer than half the alerts sent to a named owner are opened on the day they land, the volume is above what that person can absorb. Measure your own read rate for a week.

Should I use keywords or CPV codes for tender alerts?

Use both, with codes as the spine and keywords at the edges. Classification codes such as CPV in Europe and the UK, or NAICS and PSC in US federal contracting, are applied by the buyer to every notice, so they give a floor of coverage that does not depend on wording. Keywords work inside a broad code.

Can I filter tender alerts by contract value?

On most portals yes, but check what the filter does with notices that leave the value field empty: a minimum value filter commonly excludes blanks as well as small contracts. Those are notices where the buyer published no estimate, not small work.

Why does the same tender reach me twice?

Above-threshold procurements are often published on a national portal and on a wider service as well, so one procurement produces two notices with different reference numbers and titles. Deduplicate on the buyer's own procurement reference rather than the title.